[FIXED] How to tune tesseract for identifying number plate of a car more accurately?

Issue

I have a code to detect and identify the car number plate and convert the image into text using tesseract. I am using openCV to localise the number plate. The problem that I am facing is that tesseract is not accurately identifying the number. Is there any way I can improve the tesseract performance?

My code (which I downloaded from Internet) is:

import numpy as np
import cv2
# from copy import deepcopy
from PIL import Image
import pytesseract as tess

# plate = 0
def preprocess(img):
    # print ('preprocessing image')
    # cv2.imshow("Input", img)
    imgBlurred = cv2.GaussianBlur(img, (5, 5), 0)
    gray = cv2.cvtColor(imgBlurred, cv2.COLOR_BGR2GRAY)

    sobelx = cv2.Sobel(gray, cv2.CV_8U, 1, 0, ksize=3)
    cv2.imshow("Sobel",sobelx)
    cv2.waitKey(0)
    ret2, threshold_img = cv2.threshold(sobelx, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
    cv2.imshow("Threshold",threshold_img)
    cv2.waitKey(0)
    return threshold_img


def cleanPlate(plate):
    # print ("CLEANING PLATE. . .")
    gray = cv2.cvtColor(plate, cv2.COLOR_BGR2GRAY)
    kernel = cv2.getStructuringElement(cv2.MORPH_CROSS, (3, 3))
    thresh= cv2.dilate(gray, kernel, iterations=1)

    _, thresh = cv2.threshold(gray, 150, 255, cv2.THRESH_BINARY)
    im1, contours, hierarchy = cv2.findContours(thresh.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)

    if contours:
        areas = [cv2.contourArea(c) for c in contours]
        max_index = np.argmax(areas)

        max_cnt = contours[max_index]
        max_cntArea = areas[max_index]
        x, y, w, h = cv2.boundingRect(max_cnt)

        if not ratioCheck(max_cntArea, w, h):
            return plate, None

        cleaned_final = thresh[y:y + h, x:x + w]
        # cv2.imshow("Function Test",cleaned_final)
        return cleaned_final, [x, y, w, h]

    else:
        return plate, None


def extract_contours(threshold_img):
    # print ('extracting contours')
    element = cv2.getStructuringElement(shape=cv2.MORPH_RECT, ksize=(17, 3))
    morph_img_threshold = threshold_img.copy()
    cv2.morphologyEx(src=threshold_img, op=cv2.MORPH_CLOSE, kernel=element, dst=morph_img_threshold)
    cv2.imshow("Morphed", morph_img_threshold)
    cv2.waitKey(0)

    im2, contours, hierarchy = cv2.findContours(morph_img_threshold, mode=cv2.RETR_EXTERNAL,
                                                method=cv2.CHAIN_APPROX_NONE)
    return contours


def ratioCheck(area, width, height):
    # print ('checking ratio')
    ratio = float(width) / float(height)
    if ratio < 1:
        ratio = 1 / ratio

    aspect = 4.7272
    min = 15 * aspect * 15  # minimum area
    max = 125 * aspect * 125  # maximum area

    rmin = 3
    rmax = 6

    if (area < min or area > max) or (ratio < rmin or ratio > rmax):
        return False
    return True


def isMaxWhite(plate):
    # print ('is Max white')
    avg = np.mean(plate)
    if (avg >= 115):
        return True
    else:
        return False


def validateRotationAndRatio(rect):
    # print( 'validate the rotation and ratio')
    (x, y), (width, height), rect_angle = rect

    if (width > height):
        angle = -rect_angle
    else:
        angle = 90 + rect_angle

    if angle > 15:
        return False

    if height == 0 or width == 0:
        return False

    area = height * width
    if not ratioCheck(area, width, height):
        return False
    else:
        return True


def cleanAndRead(img, contours):
    # print ('clean and read')
    # count=0
    for i, cnt in enumerate(contours):
        min_rect = cv2.minAreaRect(cnt)

        if validateRotationAndRatio(min_rect):

            x, y, w, h = cv2.boundingRect(cnt)
            plate_img = img[y:y + h, x:x + w]

            if (isMaxWhite(plate_img)):
                # count+=1
                clean_plate, rect = cleanPlate(plate_img)

                if rect:
                    x1, y1, w1, h1 = rect
                    x, y, w, h = x + x1, y + y1, w1, h1
                    cv2.imshow("Cleaned Plate", clean_plate)
                    cv2.waitKey(0)
                    plate_im = Image.fromarray(clean_plate)
                    plate_im.save('donald1.png')
                    text = tess.image_to_string(plate_im, lang='eng')
                    # print text
                    img = cv2.rectangle(img, (x, y), (x + w, y + h), (0, 255, 0), 2)

                    cv2.imshow("Detected Plate", img)
                    cv2.waitKey(0)
                    return text





numberplate = 0


img = cv2.imread("car_number_plate.jpg")

threshold_img = preprocess(img)

contours = extract_contours(threshold_img)

    # if len(contours)!=0:
    # print len(contours) #Test
    # cv2.drawContours(img, contours, -1, (0,255,0), 1)
    # cv2.imshow("Contours",img)
    # cv2.waitKey(0)


plate = cleanAndRead(img, contours)
print ('plate information: ', plate)

If my number plate is: MH01AV8866

It will be recognised as MH01AY8866

Any suggestion will be appreciated. Let me know if any other information is required too.

Solution

You are using tesseract as a general model for your problem you can tune your model for that you need to generate synthetic data for your number plates with this

https://github.com/Belval/TextRecognitionDataGenerator

and then you can tune your model using the steps provided

https://github.com/tesseract-ocr/tesseract/wiki/TrainingTesseract-4.00---Finetune

https://github.com/tesseract-ocr/tesseract/wiki/TrainingTesseract-4.00

I've tuned the tesseract on synthetic data and it works like a charm, tried CNN models and tesseract both and tesseract trains better with lesser data and gives better performance.

Answered By - Yash Kumar Atri